Python framework to record and validate data pre-processing steps
in Data Science and AI workflows. Still in
pip install python-dataflow
If you have some raw data you often apply some functions that change your data to fit a specific problem. The problem is that there is no simple tool that can record and validate all the changes that happened to your data while you were working with it. It becomes a bigger problem when you have lots of pre-processing steps or work in a team with other developers.
By using Dataflow, you can easily wrap your functions with decorators and record/validate all the changes. You can think of Dataflow as a migration tool, but for data pipelines.
Without Dataflow, your code can look like this:
raw_data = [1, 2, 3, ...]
unique_data = list(set(raw_data))
sorted_data = sorted(unique_data)
final_data = [
value for value in sorted_data if value >= 5
]
print(final_data) # [2, 3, 4, 5]It is a simple example, buy you can imagine your Torch, Tensorflow, Pandas or Numpy pre-processing here. However, there are some problems:
- Are we sure the data is correctly processed?
- What changes were made and in what order?
- Can we replicate the changes easily?
These problems have haunted many, but you can solve them easily with Dataflow:
from dataflow.mutations import mutation
from dataflow.validations.pipelines.after import after
from dataflow.validations.pipelines.before import before
@mutation() # Mutation is a step in your pipeline
@before(lambda data: len(data) > 0, hint="Data is empty") # Data is not empty before we run the mutation
@after(lambda data: len(data) > 5, hint="Data size is less than 5") # Data has enough entries after the mutation
def remove_duplicates(data: list[float]) -> list[float]:
return list(set(data))
@mutation() # Another mutation without any validations
def sort(data: list[float], reverse: bool = False) -> list[float]:
return sorted(data, reverse=reverse)
@mutation() # Mutation with a simple validation
@after(not_lambda_validation)
def remove_under_threshold(data: list[float], threshold: float) -> list[float]:
return [value for value in data if value >= threshold]
mutation_pipeline = sort() >> remove_duplicates() >> remove_under_threshold(threshold=2)
final_data = mutation_pipeline(data=[1, 2, 3, 4, 5] * 100) # final data = [2, 3, 4, 5]
print(mutation_pipeline) # sort() >> remove_duplicates() >> remove_under_threshold(threshold=2)As you can see, we don't change much in the data pre-processing itself, but only add some decorators and use some new syntax to validate that the changes are correct. Dataflow enables users to easily find pitfalls and issues in their pipelines.
- Friendly -> you can use Dataflow with any environment and run mutations in various data-sources. Mutations and validations can be extended to work anywhere.
- Easy to master -> you only need to learn a few basic commands and don't have to rewrite your whole codebase to fit this framework.
- Extendable -> write your custom mutations, validations, pipelines without worrying about breaking something internal. Dataflow has very few rules on how to write the code.
MIT